Scientists from AI4FLOOD, a Helmholtz AI project that received funding in our 2019 call, have developed a deep learning algorithm for flood detection that won third place in the NASA-ETCI competition.
Flood events occur all over the world and can have devastating consequences for people, ecosystems and economies: a single major flood can cause billions of dollars in damage. The hazards of using ground-based equipment in floodplains and limited physical access to flooded areas make it difficult to acquire information about the extent of flooding on the ground. That’s why capturing floods and the extent of flooding remotely and accurately is a great help in responding to these destructive events and in mitigating their effects.
To advance the use of machine learning (ML) and artificial intelligence (AI) in remote flood detection, NASA’s Interagency Implementation and Advanced Concepts Team (IMPACT) has organized the Emerging Techniques in Computational Intelligence (ETCI) 2021 Competition on Flood Detection. The Institute of Electrical and Electronics Engineers (IEEE) Geoscience and Remote Sensing Society (GRSS) Earth Science Informatics Technical Committee (ESI TC) joined IMPACT in coordinating this effort. The challenge involved a supervised learning task in which participants developed algorithms to identify flood pixels after training their algorithm on a training set of synthetic aperture radar (SAR) images.
GFZ scientists used this opportunity to train a deep-learning algorithm for flood detection as part of the 2019 Helmholtz AI funded project AI4Flood led by Prof. Dr. Mahdi Motagh from the German Research Centre for Geosciences (GFZ) and Dr. Sandro Martinis from the German Aerospace Center (DLR). Specifically, Shagun Garg from GFZ Potsdam proposed a convolutional neural network (CNN) based on the Unet architecture with a backbone of EfficientNetb7, which was trained with the competition dataset. The performance of the model was then evaluated using several training, testing, and validation methods. Two evaluation methods - Intersection over Union (IOU) and F-Score - are applied to assess model performance. In the tests, the model developed by GFZ achieved an average IOU score of 75.06% and an F-Score of 74.98%. This result placed it among the top three algorithms in the NASA-ETCI competition. Further information on the competition can be found here: https://nasa-impact.github.io/etci2021/